[Wannier] dis_project ZGESVD error

[Nicola Marzari]

Dear Brad,


a guess is that your random projections, projected on the kohn-sham
manifold, end up having a couple of them that are almost linearly
dependent.

Can you choose instead a regular mesh of centers for the projections,
or maybe something slightly more physical?

If your system is large, you could try the gamma sampling only -
that typically does not need projections, and you could use the results 
as centers for the next set of projections?

		nic


On 12/7/11 12:11 PM, Jonathan Yates wrote:
>
> On 6 Dec 2011, at 19:25, Brad Malone wrote:
>
>> Hi, I am using wannier in library mode for calculations on a system as a function of pressure. I am able to successfully wannierize the system for 20,25,30, and 40 GPa with essentially no change in the p.win file except for a slight change to the frozen window range (to account for the increase in the bandwidth as a function of pressure).
>>
>> However, when I try to do it for 50 GPa, for the same system and parameters, I get the error:
>>
>>
>>                    Unitarised projection of Wannier functions
>>                    ------------------------------------------
>>    A_mn =<psi_m|g_n>  -->  S = A.A^+ -->  U = S^-1/2.A
>>    In dis_project...  ERROR: IN ZGESVD IN dis_project
>>   K-POINT NKP=            1  INFO=            7
>> Exiting.......
>> dis_project: problem in ZGESVD 1
>>
>> I was originally using 'random' projections, since that is what has worked for me for the previous pressures. When I saw this message I tried fiddling with the projections, the number of wannier functions, and the number of input bands, but nothing changed (except changing the number of wannier functions changes the value for INFO).
>
> Brad,
>
> The error means that that singular value decomposition has failed to converge. It's being used to compute the initial guess for the unitary rotation.
>
>   I've never encountered this error before, and from your description it is not at all obvious what might cause it. Perhaps it is worth trying a different set of math libraries. (eg if you are using mkl, try the netlib lapack).
>
>   Otherwise, if no-one else has a solution, you could send the files (win, eig, amn, mmn) to me.
>
>   Jonathan
>
>
>
>


-- 

----------------------------------------------------------------------
Prof Nicola Marzari, Chair of Theory and Simulation of Materials, EPFL